This report summarizes the findings, recommendations, and process by the VIS Area Curation Committee (ACC) regarding the areas used for paper submissions to IEEE VIS 2024. We use submission information from VIS 2024 to analyze the impact of the topics to the area model.
Given the information we have access to, the areas seem relatively balanced, and we do not make recommendations for any changes, creation or merging, at this time.
The full data and source code to rebuild this project are available here.
Last edited: 2024-10-04.
%load_ext autoreload
%autoreload 2
def in_notebook():
try:
from IPython import get_ipython
if 'IPKernelApp' not in get_ipython().config: # pragma: no cover
return False
except ImportError:
return False
except AttributeError:
return False
return True
IN_NOTEBOOK = in_notebook()min_df = 25
max_df = 0.1
max_words = 100000
vocab_sample = 250000
num_dims = 50
filt_min_score = 4
n_neighbors = 10
llm_family = "specter2"
# llm_family = "all-MiniLM-L6-v2"import pandas as pd # noqa
df = pd.read_csv('../data/papers_all.csv', index_col=None)
df.abstract = df.abstract.fillna('')
df.keywords = df.keywords.fillna('')
# print(df.columns)
# prints: Index(['abstract', 'title', 'keywords', 'Area', 'year'], dtype='object')
df['text'] = df.abstract + ' ' \
+ df.keywords + ' ' \
+ df.title
# df = df[["Year", "Title", "Abstract", "AuthorKeywords", "text"]]
docs = list(df.text)
# df.head()We pass the 2021-2024 abstract texts to the Bertopic library and it computes topics magically. These topics seem representative of the main trends in visualization during the period.
from bertopic import BERTopic
from sentence_transformers import SentenceTransformer
from sklearn.feature_extraction.text import CountVectorizer
from bertopic.representation import MaximalMarginalRelevance
from umap import UMAP
vectorizer_model = CountVectorizer(ngram_range=(1, 3), stop_words="english")
representation_model = MaximalMarginalRelevance(diversity=0.3)
sentence_model = SentenceTransformer('sentence-transformers/allenai-specter')
embeddings = sentence_model.encode(docs, show_progress_bar=IN_NOTEBOOK)
topic_model = BERTopic(representation_model=representation_model)
topics, probs = topic_model.fit_transform(docs, embeddings)
topic_model.update_topics(docs, vectorizer_model=vectorizer_model)
with pd.option_context('display.max_rows', None, 'display.max_columns', None, 'display.max_colwidth', 1000, 'display.width', 200):
print(topic_model.get_topic_info()[["Topic", "Count", "Representation"]])
# df_model_last = topic_model.get_topic_info()
# df_model_last.to_csv("topics-all.csv", index=False)/home/fekete/miniforge3/envs/acc-nlp/lib/python3.10/site-packages/numba/np/ufunc/parallel.py:371: NumbaWarning: The TBB threading layer requires TBB version 2021 update 6 or later i.e., TBB_INTERFACE_VERSION >= 12060. Found TBB_INTERFACE_VERSION = 12050. The TBB threading layer is disabled.
warnings.warn(problem) Topic Count Representation
0 -1 171 [data, visualization, visual, design, analysis, visualizations, users, approach, study, model]
1 0 52 [data, rendering, fields, volume, field, visualization, merge, topological, approach, scientific]
2 1 42 [visualization, visualizations, uncertainty, design, data, causal, visual, trust, decision, participants]
3 2 37 [network, layout, graph, visualization, data, networks, visual, diagrams, graphs, set]
4 3 29 [urban, time, series, data, visual, time series, analysis, analytics, visual analytics, visualization]
5 4 23 [data, interaction, immersive, reality, design, ar, spatial, situated, visualizations, physical]
6 5 20 [authoring, design, visualization, data, visualization authoring, dashboard, pictorial, interactive, comics, visualizations]
7 6 19 [dimensionality, reduction, dimensionality reduction, embeddings, embedding, highdimensional, dr, cluster, data, class]
8 7 17 [llms, language, models, language models, model, large language, large language models, large, visualization, learning]
9 8 17 [textures, line, colors, associations, perceptual, visual, charts, visualization, visualizations, concepts]
10 9 17 [data, video, stories, data stories, videos, animation, data videos, design, story, narrative]
11 10 16 [sports, player, tactics, basketball, players, badminton, game, tactic, visualizations, videos]
12 11 13 [model, neural, learning, concepts, deep, neural networks, networks, visual, object, deep learning]
13 12 12 [patient, data, cancer, patients, clinical, cells, treatment, disease, visual, progression]from scipy.cluster import hierarchy as sch
import numpy as np
linkage_function = lambda x: sch.linkage(x, 'ward', optimal_ordering=True)
hierarchical_topics = topic_model.hierarchical_topics(docs, linkage_function=linkage_function)
topic_model.visualize_hierarchy(hierarchical_topics=hierarchical_topics) 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████| 12/12 [00:00<00:00, 192.89it/s]These topics can be visualized over a map.
if IN_NOTEBOOK:
reduced_embeddings = UMAP(n_neighbors=10, n_components=2, min_dist=0.0, metric='cosine').fit_transform(embeddings)
vis = topic_model.visualize_documents(df.title, reduced_embeddings=reduced_embeddings, hide_annotations=True, title='<b>Documents and Topics 2021-2024</b>')
# vis.write_html("topics-all.html")
else:
vis = None
visTo find out if some topics are over- or under-represented, we visualize the topics per Area. The largest area, “Application”, contains a relatively balanced mix of topics; there is no clear thematic way of splitting it.
topics_per_class = topic_model.topics_per_class(docs, classes=list(df.Area))
fig = topic_model.visualize_topics_per_class(topics_per_class, top_n_topics=10)
fig.for_each_trace(lambda bar : bar.update(visible=True))